Graff‐Radford, Jonathan

AbstractBackgroundResearch subjects can potentially be re‐identified from de‐identified MRI, CT, and PET brain scans with up to 98% accuracy using Microsoft Azure’s cloud‐based commercial facial recognition software. This showed the need to “de‐face” publicly shared research brain scans. Subsequently, Microsoft has begun restricting its face recognition services, intending to prevent misuse. This study tests a variety of popular open‐source computer vision and facial recognition software packages to measure the re‐identification risk with only unrestricted free software.MethodThis study used brain MRI from Siemens 3T scanners and face photographs from 182 participants (ages 30‐90+) in the Mayo Clinic Study of Aging. Each participant had photos taken with an iPad, which were cropped loosely around the head and converted to black and white. A total of 81 renders (2D reconstruction images; Figure 1) were made from each 3D T2‐FLAIR MRI, simulating various camera and lighting positions. These face renders approximate black and white photos, allowing use with software packages developed for matching faces/features between photographs. Open‐source Python packages (OpenCV, DeepFace, face_recognition, scikit‐learn) and their included pre‐trained models (designed for photographs) were used to match participant photographs to the labelled library of MRI‐based face renders. A correct match was recorded when the top‐scoring render or set of renders automatically chosen by the algorithm was from the MRI of the correct participant.ResultFeature matching with ORB and SWIFT (both from OpenCV) produced no correct matches, despite a prior publication reporting moderate success. The best match rates (Figure 2) were 58‐59% accuracy using face_recognition or DeepFace packages with support vector machines (SVMs). Without SVMs, these packages performed modestly worse (51%‐55%).ConclusionWhile the highest match rates obtained by widely available open‐source packages (roughly 60%) are much lower than those of a limited‐access commercial product (Microsoft Azure, with 98%), they show that actors with only access to no‐cost readily available software could still re‐identify a research subject from a brain scan most of the time. This further highlights the need for image refacing software, such as mri_reface, to protect the privacy and identifies of research volunteers.

AbstractBackgroundSeveral programs are available for automatically “de‐facing” brain images. These are designed to prevent the potential use of face imagery to re‐identify research participants. Previous works have compared available software and mri_reface is frequently among the top methods. The “Minimally Invasive DeFacing of MRI images” (MiDeFace) program was more recently added to FreeSurfer, but minimal validation of MiDeFace has been published.MethodWe ran mri_reface 0.3.3 and MiDeFace 7.3.2 on two datasets: A) 207 participants with face photographs and 3D T1‐weighted and 3D T2‐weighted‐FLAIR MRI from the Mayo Clinic Study of Aging (all unimpaired, Siemens); and B) 300 participants from ADNI3 with 3D T1‐weighted MRI (50 unimpaired + 50 with clinical AD, from each of 3 vendors). For dataset A, we tested the rate at which Microsoft Azure face recognition matched each participant’s photographs with the correct MRI, comparing unmodified images, those from mri_reface, and those from MiDeFace. For dataset B, we compared measured regional brain volume and cortical thickness from original vs. de‐faced images with SPM12, FreeSurfer 7.4.1, and the FSL‐UK Biobank pipeline. For both datasets, images were visually rated to assess removal of the face and retention of brain tissue.ResultAutomated face recognition match rates were lower with mri_reface (17/207 8.2%) than with MiDeFace (24/207 11.6%), indicating stronger identity protection with mri_reface, but both were greatly reduced vs. unmodified images (203/207 98.1%). Example images with both programs are shown in Figure 1. From visual ratings (Table 1), rates of complete success (modification of the entire face while retaining the entire brain) were much higher with mri_reface, but the great majority of failures from MiDeFace were subtle and occurred in relatively less recognizable face regions. The effects of both programs on downstream imaging measurements (Figure 2) were small and very comparable, but MiDeFace was favored mildly for SPM12 and for FSL, while mri_reface was more favored for FreeSurfer.ConclusionAltogether, mri_reface more reliably removed the complete face, leading to smaller rates of potential re‐identification (8.2% vs 11.6%). Effects on downstream image measurements were small and comparable with both methods.

AbstractBackgroundThe Mayo Clinic Study of Aging (MCSA) is a longitudinal, population‐based study of residents of Olmsted County, Minnesota, USA. MCSA is releasing de‐identified clinical and imaging data on GAAIN.org to benefit the research community.MethodWe included longitudinal clinical data from all participants of the Mayo Clinic Study of Aging 30‐90 years of age (average 70.0y) through the first 15 years of the study. Clinical data includes age, sex, self‐reported race/ethnicity, APoE4 allele status, cognitive impairment status (normal, MCI, dementia, or other), height, weight, blood pressure, multi‐domain neuropsychological scores, neuroimaging summary measurements, medications, cardiovascular risk factors, social activity, sleep, mood/anxiety, functional measures, and more. For our initial release of imaging data, we selected the first imaging timepoint acquired on 3T GE MRI scanners and amyloid (PiB) PET during the same visit, from all available imaging participants 50‐90 years of age (average 70.7y). Currently, we included 3D T1‐weighted (MPRAGE) MRI (1x1x1.2mm3), T2‐weighted FLAIR MRI (0.85x0.85x3.6mm3), diffusion MRI (2.7mm resolution with 41 directions at b=1000 and 5 b=0), and numeric cortical global SUVR and Centiloid values from PiB PET scans at 40‐60 minutes post‐injection. All images are released in DICOM format and thoroughly de‐identified using CTP while ensuring preservation of necessary fields for analyses. T1‐weighted and T2‐FLAIR MRI were also de‐faced (identifiable face imagery was removed and replaced with an average) using mri_reface version 0.3.3 to further protect participants’ privacy, and all de‐faced images were visually inspected by a trained data scientist for quality assurance.ResultThe first phase of our public data release includes clinical data from 5925 unique participants from the Mayo Clinic Study of Aging (MCSA), each ranging from 1‐12 visits at roughly 15‐month intervals, and three MRI sequences (T1‐w, T2‐w‐FLAIR, and diffusion) from 1802 unique participants. Demographic summaries are given in Table 1.ConclusionThis large, richly characterized, de‐identified clinical and imaging dataset is available to the research community at https://www.gaaindata.org/partner/MCSA, and it is suitable for a wide range of research uses. Additional data will be released over time, including de‐faced amyloid PET images.

AbstractBackgroundWhite matter (WM) damage is seen with neurodegenerative and cerebrovascular pathologies and contributes to cognitive dysfunction. We hypothesized that there are multiple disease progression patterns in WM microstructural changes related to aging and dementia, and these can be identified using SuStaIn (data‐driven clustering algorithm for disease subtype and stage discovery) on multishell diffusion MRI (NODDI) data. NODDI provides information on cellular tissue architecture – neurite density index (NDI) measures packing density of axons and dendrites, and free water fraction (FWF) measures unrestricted water, e.g., interstitial and cerebrospinal fluid.MethodWe identified 1136 participants (mean age=70 years, 51% male, 40% Aß‐positive, 19% MCI/AD dementia) from the Mayo ADRC and Mayo Clinic Study of Aging. We analyzed NDI and FWF in 55 WM regions. We selected a subset of WM features that monotonically changed with clinical disease progression and then applied the Z‐score SuStaIn algorithm.ResultNDI in seven regions and FWF in six regions survived selection criteria (Figure 1). Noteworthy, the NDI regions were predominantly located in the temporal lobe, while FWF regions were primarily distributed in the frontal and subcortical regions. The SuStaIn algorithm discovered two stable clusters: the NDI‐First (Temporal) and FWF‐First (Frontal) groups (Figure 2). For the NDI‐First group, temporal lobe NDI changed early followed by frontal FWF. The FWF‐First group had the reverse order. There was clear phase separation based on which measure changed first. Stages of the NDI‐First group correlated better with meta‐ROI Tau‐PET SUVR and with MMSE than the FWF‐First group (p<0.01), suggesting that NDI‐First may be indicative of a primarily AD pathway.ConclusionGiven NDI changes are likely degenerative and FWF are suggested to be other processes such as neurodegeneration or inflammatory, the observed changes in the temporal and frontal regions may be due to distinct etiologies, i.e., neurodegenerative in the temporal lobes vs. vascular dysfunction often observed in the frontal lobes. The SuStaIn algorithm showed promise in discovering disease subtypes of WM microstructural changes. Further work is needed to understand the mechanistic basis of these observed changes and integrate WM changes for differential etiology identification and disease staging.

AbstractBackgroundImaging and plasma markers are used as key indicators of disease for Alzheimer’s disease (AD) but their usefulness in predicting regional tau pathology is relatively understudied. Our objective was to construct predictive models for regional tau pathology measured on postmortem brain tissue using multiple ante‐mortem AD biomarkers. We focused on hippocampal and parietal regions that were immunostained with AT8 and 2E9 that reflect early and advanced aspects of tangle maturity, respectively.MethodsUsing autopsy data from 63 participants enrolled in the Mayo Clinic Study of Aging (MCSA) where the last clinical visit is within three years of death, we analyzed three ante‐mortem plasma markers ‐ phosphorylated tau protein at the threonine 181 residue, glial fibrillary acidic protein, total ß‐amyloid Aß42 relative to Aß40 ratio (Quanterix Simoa HD‐X Analyzer), and three regional MRI markers computed using SPM12 ‐ parietal thickness, parietal volume relative to total intracranial volume ratio, and hippocampal volume adjusted for total intracranial volume. Based on eight separate cross‐sectional linear regression models weighted by the inverse of time from last visit to death, we estimated the multiple and partial R2 predicting post‐mortem pathology measures, including AT8 and 2E9 in both the hippocampal and parietal regions. All models were adjusted for age, sex, and cardiovascular and metabolic conditions score.ResultsThe MRI models (R2 of 0.26‐0.53) had greater predictive power compared to the plasma models (R2 of 0.17‐0.26). The best multiple R2 was 0.53 for prediction of hippocampal AT8 using antemortem hippocampal volume adjusted for total intracranial volume followed by R2 of 0.43 for prediction of hippocampal 2E9 using hippocampal volume adjusted for total intracranial volume. Though the model prediction of parietal AT8 and 2E9 measurements was low (R2 of 0.21‐0.26), parietal thickness exhibited better predictive performance than parietal volume relative to total intracranial volume ratio.ConclusionsThe data we presented here may be utilized to validate biomarkers using regional pathology. Our results support that regional MRI is sensitive to tau pathology. Newer MRI methods should be tested to examine the specificity of newer quantitative measurements to tau pathology.

AbstractBackgroundQuantifying white matter using diffusion MRI (dMRI) has been proposed for measuring early microstructural tissue changes due to cerebral small vessel disease and aid in quantifying vascular contributions to cognitive impairment and dementia (VCID). Our goal was to compare the usefulness of longitudinal white matter changes in the commonly available diffusion MRI measures for VCID prevention trials.MethodWe included 718 participants over 50 years of age (mean age: 71.1(9.6) years) from the Mayo Clinic Study of Aging, a population‐based sample, with at least two dMRI scans and structural imaging. We computed the commonly available single‐shell dMRI measures (fractional anisotropy, mean diffusivity, and two MarkVCID measures ‐ free water and peak‐width skeletonized mean diffusivity [PSMD]) at each time point. We tested for voxel‐wise associations between dMRI markers and vascular risk measured by cardiovascular metabolic condition (CMC) and observed a region‐specific dependance across all dMRI measures. Using both global and regional dMRI measures, we (i) examined the longitudinal association of dMRI measures with cognition and (ii) computed sample size estimates for a hypothetical clinical trial. We also included white matter hyperintensities (WMH) and our previously proposed composite vascular white matter score (combination of WMH and fractional anisotropy of the genu) as a comparison.ResultVascular risk was associated with all single shell dMRI measures in the genu of the corpus callosum, which we included as a regional dMRI marker for comparison (Figure 1). All dMRI markers correlated with cognitive performance longitudinally (Table 1) and had comparable sample size estimates required for hypothetical VCID clinical trials (Figure 2). Further, global free water and the composite vascular white matter score had the smallest sample size estimates.ConclusionAll commonly used dMRI markers had significant frontal lobe changes due to vascular risk. Both global and regional corpus callosum dMRI markers were sensitive to longitudinal cognitive decline, suggesting their utility in measuring the slowing down of VCID. The composite vascular white matter score, global free water, and WMH show promise as VCID biomarkers. Further work is needed to validate these markers on multiple populations.

AbstractBackgroundThere has been a recent proliferation of various quantities based on single shell Diffusion Tensor Imaging (DTI) for capturing overall cerebrovascular health, especially Small Vessel Disease (SVD) as a single number. The existing literature has gaps in the comparison of these measures, so we evaluated them as predictors of both current and future White Matter Hyperintensity (WMH) as a proxy of SVD.MethodUsing baseline 3T MRI examinations (T1, DTI, FLAIR scans) of 598 participants (>60 years, amyloid negative, to limit to those on the SVD pathway) from the Mayo Clinic Study of Aging, we calculated five summary DTI quantities with varying degree of complexity. We estimated their partial R2 from cross‐sectional linear regression models and computed semi‐partial R2 from mixed models predicting WMH volume as a percentage of Total Intracranial Volume (TIV). All models were adjusted for age, sex, and scanner model, and the WMH and based were log‐transformed before fitting. We also conducted subset analyses with high (WMH+) vs. low (WMH‐) baseline WMH load (cut at 1.62%) to compare the performance of the measures in early vs. widespread WMH.ResultAll MD‐based quantities computed on the whole brain, including FWF, performed similarly to each other, but Genu FA underperformed at explaining WMH variance (Table 1). PSMD (skeletonized or atlas based) was more sensitive in WMH+ but dramatically less so in WMH‐. In WMH‐ and overall cohort, FWF and JHU MD (which is a simple averaging of MD in the JHU atlas) performed well.ConclusionAll MD‐based measures performed similarly to each other in the whole dataset but dichotomizing by WMH load showed interesting differences. PSMD measurements are driven by the extreme end (95th percentile) of the MD distribution and therefore more sensitive in WMH+. Genu FA, a regional measure, was not useful for predicting whole brain WMH change. We noticed some variation in how the quantities are affected by scanner model (GE Signa HDxt and Discovery 750) which needs broader investigation. Simply averaging MD over WM avoids additional processing and is more sensitive to WM degeneration that is not (yet) severe.

AbstractBackgroundChronic medical conditions increase dementia risk, and the mechanism may be captured computationally by measures of accelerated brain aging. We hypothesized that quantifying the impact of chronic conditions on GM and WM aging will aid in measuring the differential impact of each condition on brain aging and shed light on the mechanisms through which they increase dementia risk. We specifically focused on microstructural brain changes with Neurite Orientation Dispersion and Density Imaging (NODDI) based on advanced diffusion MRI ‐ neurite density index (NDI), free water fraction (FWF), and orientation dispersion index (ODI).MethodWe identified 1123 participants [913 cognitively unimpaired (CU), 114 with MCI, 96 with AD] from the Mayo Clinic Study of Aging (MCSA) and Mayo ADRC with multi‐shell diffusion MRI. MCSA participants had chronic medical conditions’ diagnosis (based on US Dept of HHS) extracted from the electronic health care records. Using the BorutaSHAP algorithm on NDI, ODI, and ISOVF features from GM and WM separately, we constructed two distinct brain age models for GM and WM. First, we evaluated the usefulness of the predicted age gap (PAG), which is the difference between predicted age by model and chronological age, for both WM and GM NODDI models in CU/MCI/AD. We then developed separate GM and WM models for each condition using samples from condition‐negative participants to compare PAG of GM and WM for that condition.ResultThere was greater PAG in GM (avg = ) than WM (avg = 9.15 yrs) in MCI/AD participants (Fig. 1A). Each condition shifted the GM PAG by 0.10 to 3.56 years and WM PAG by 0.38 to 3.72 years with stroke and congestive heart failure showing the greater PAG overall for GM and WM PAD respectively (Fig. 2). COPD, stroke, chronic kidney disease, hypertension had comparatively greater GM PAG than WM PAG; while hyperlipidemia, congestive heart failure and coronary artery disease had greater WM PAG than GM PAG (Fig. 1B).ConclusionChronic conditions are associated with increased GM and WM Brain aging models can aid in quantifying the impact of chronic conditions on dementia risk and understanding mechanisms.

AbstractBackgroundThere is a strong link between tau and progression of Alzheimer’s disease (AD), necessitating an understanding of tau spreading mechanisms. Prior research, predominantly in typical AD, suggested that tau propagates from epicenters (regions with earliest tau) to functionally connected regions. However, given the constrained spatial heterogeneity of tau in typical AD, validating this connectivity‐based tau spreading model in AD variants with distinct tau deposition patterns is crucial.MethodWe included 269 amyloid‐ß‐positive (PET/CSF) individuals with clinically diagnosed atypical AD (113 posterior cortical atrophy, PCA‐AD; 83 logopenic variant primary progressive aphasia, lvPPA‐AD; 33 behavioural variant AD, bvAD; 40 corticobasal syndrome, CBS‐AD) and 68 with typical AD from 12 international cohorts, who underwent tau‐PET (54% [18F]AV1451/[18F]flortaucipir/Tauvid, 27% [18F]MK6240, 19% [18F]PI2620). Using Gaussian mixture modeling including amyloid‐ß‐negative controls, cross‐sectional tau‐PET standardized uptake value ratios within Schaefer‐200 atlas regions were transformed to tau positivity probabilities. Tau epicenters were defined as the 5% regions with highest tau positivity probabilities. For each variant, the association between functional connectivity‐based distance (using the 30% strongest positive region‐to‐region connections of a group‐average connectivity matrix from ADNI elderly controls) and tau‐PET covariance (group‐average correlation per region pair) was assessed through linear regression, adjusting for age, sex, site, and Euclidean distance. Regions were categorized based on functional proximity to the epicenter (quartiles 1‐4) and tau positivity probabilities were assessed accordingly.ResultTau positivity probabilities matched clinical variants, with a posterior pattern in PCA‐AD, left‐hemispheric dominant pattern in lvPPA‐AD, widespread pattern in bvAD, sensorimotor cortex involvement in CBS‐AD, and temporo‐parietal predominance in typical AD (Fig.1). In line with this, tau epicenters were highly heterogeneous across variants (Fig.1). In all variants, greater tau‐PET covariance was associated with shorter functional connectivity‐based distance (Fig.2). We observed that regions in closer functional proximity to the epicenter exhibited higher tau positivity probabilities than regions functionally further away (p<0.05, Fig.3).ConclusionThis multi‐center study shows that the brain’s functional architecture serves as a universal predictor of tau spreading in AD. Since tau is a key driver of neurodegeneration and cognitive decline in AD, this finding holds potential for personalized medicine and defining participant‐specific endpoints in clinical trials.

Abstract Background Many proposed clinical decision support systems (CDSS) require multiple disparate data elements as input, which makes implementation difficult, and furthermore have a black‐box nature leading to low interpretability. Fluorodeoxyglucose Positron Emission Tomography (FDG‐PET) is an established modality for the diagnosis of dementia, and a CDSS that uses only an FDG‐PET image to produce a reliable and understandable result would ease both of these challenges to clinical application. Method A deep variational autoencoder (VAE) was used to extract a latent representation of each image through prior training from FDG‐PET brain images (n=2000). This unsupervised VAE has a novel graph convolutional architecture that makes it applicable to masked template space images. A logistic regression model was used to classify each image. A parametric study of the latent space revealed the imaging features that were used in the logistic regression model to differentiate each class. A separate dataset of participants labeled with their clinical diagnosis (n=1239) was used to assess the logistic regression model’s diagnostic accuracy of cognitively unimpaired (CU) (n=679) and between common dementia subtypes: Alzheimer’s disease (AD) (n=310), Lewy body dementia (DLB) (n=151), and behavioral variant frontotemporal dementia (bvFTD) (n=99). The logistic regression classifier was evaluated using receiver‐operator characteristic (ROC‐AUC) curves. Result ROC‐AUC curves for CU and each dementia subtype are illustrated in Figure 1a. A k nearest neighbors model was used to develop a graphical representation of the latent space where nodes are images and edges are drawn between nearest neighbors, illustrated in Figure 1b. The graph of the latent space demonstrates a comparison between dementia subtypes and CU images. Conclusion In this study, we developed a proof‐of‐concept VAE that can achieve differential diagnosis with 84.8% balanced accuracy. The model architecture provides visual interpretability, and shows that the hippocampi, parietal lobes, and cerebellum, were useful for distinguishing AD, the frontal lobe, posterior cingulate cortex, and hippocampi were useful for distinguishing DLB, and prefrontal pole, subcortical structures, and occipital lobe were useful for distinguishing bvFTD. These regions of metabolism align with clinical pathology and metabolic patterns.